Northstar / July 2026

Sibyl

Your entire life. Every agent. One memory that answers to you.

A source-grounded product, architecture, and roadmap vision for turning Sibyl from the unusually integrated cross-agent memory system into the sovereign context runtime beneath every agent, project, relationship, archive, and chapter of a life. Humans own the evidence. Governed agents help memory become more useful without becoming its sovereign.

Audited at 8df846ea Primary-source research Practical before magical Self-hosted by default Governed agent authorship
Enter the atlas
01 / The verdict

Sibyl is closer than the roadmap admits.

The core is no longer a memory experiment. It is a credible multi-tenant context runtime with provenance, tasks, synthesis packets, usage feedback, team scopes, Open Knowledge Format export, and unusually honest evaluation machinery. The next leap is not another retrieval trick. It is turning that runtime into a trustworthy home for an entire life.

Strategic conclusion
Build the sovereign context operating system.

A provenance-complete evidence fabric that turns raw experience and agent trajectories into precise, permissioned, replayable context and reusable competence. Not a chatbot. Not a vector database. Not a note app. Not an agent framework.

Repository reality
1.0.2current version
9built-in source shapes
7fused retrieval signals
1governed context plane
The roadmap correction

Run two destinations on one substrate. Governed authorship and team-scope trust start now; exact evidence and life-scale ingest advance beside them. Full multi-writer coalescence waits for correction lineage, but the personal archive never waits for coalescence.

02 / Ground truth

The machine we already have.

This vision was checked against the running architecture and current main branch, not merely the prose. The result is encouraging and a little spicy: Sibyl has quietly outrun parts of its own roadmap.

Source truth
Raw memories, source adapters, checkpoints, dedupe receipts, skipped records, attachment metadata, and transformation behavior form a real provenance spine.
shipped
Context runtime
Exact, full-text, vector, graph, and temporal paths converge behind org-scoped services. Context uses heuristic budgets; synthesis currently produces deterministic evidence packets rather than generated narrative.
shipped
Usage loop
Exposure and citation events can influence usage-aware ranking and decay. Action and outcome valuation remain proposed; live proof and write-path scaling are unfinished.
proving
Agent authorship
The internal ontology is rich, but the agent-facing pen is narrow. MCP remember exposes 11 of 33 entity kinds; capture creates only generic, project, and task relations; nine correction actions live behind admin REST; and basis, confidence, retention, fidelity, and negative outcomes lack one governed mutation contract.
critical gap
Team control
Team scopes, sharing, membership, and backend controls landed. The web experience still presents organization membership, not the full team-memory model.
partial
Life-scale ingest
MBOX, Maildir, IMAP, file, folder, URL, text, Claude, and Codex adapters prove the abstraction. The import ledger, blob plane, UX, and scale tests are prototype-grade.
critical gap
Evaluation
QA lanes and LongMemEval V2 plumbing exist. The release still needs pinned accuracy, scored runs, load evidence, and deletion tests.
receipts due
StrongSource-preserving provenance and citations
StrongPhysical namespace isolation per organization
StrongTasks and memory inside one graph
RiskSeveral core modules exceed 2,000–3,000 lines
UnknownSingle-node scale envelope under real corpus load
Scale trap 01

The import parent row grows forever.

A source import currently accumulates raw IDs, source IDs, dedupe keys, skipped records, and source-to-memory mappings on one parent. At hundreds of thousands of items, resume repeatedly reconstructs and serializes a giant mutable object.

Scale trap 02

Email batches revisit the archive.

MBOX batches reopen and materialize keys; Maildir batches rebuild and sort the listing. Resume is ordinal-based rather than byte-indexed, turning a large archive into repeated whole-corpus work.

Truth trap 03

“Raw” mail is extracted text.

A sovereign archive must preserve exact RFC/MBOX bytes, MIME structure, labels, headers, and attachment hashes. Parsed text is a useful projection. It is not raw evidence.

Subtract before expanding

Make one obvious path through the power already here.

An internal audit of 1,732 local agent transcripts found ritual-heavy read verbs, seven importance keys, four project-scope expressions, narrow use of a 33-kind ontology, and correction machinery separated from ordinary capture. Unify the authoring and recall surfaces, collapse duplicate fields and scope resolution, and advertise a smaller coherent vocabulary without deleting the richer backend model. Transcript counts are host-conditioned evidence, not a referendum on capability.

Audited against the repository overview, Northstar, the post-1.0 roadmap, source adapters, import state, retrieval services, usage services, web surfaces, tests, and git history at 8df846ea. The linked SIBYL_1_1_IMPLEMENTATION_PLAN.md does not exist in this tree.

03 / The field

The systems surveyed here each own a piece. None spans this whole.

The frontier is fragmenting into coding provenance, agent memory, temporal graphs, local capture, portable data, and platform-managed context. Sibyl’s opportunity is to compose those truths into one governed life substrate without pretending one index solves every problem.

Coding provenance

Entire

Git-linked checkpoints preserve prompts, transcripts, tool calls, file changes, tokens, and attribution. They support review and search; resume and rewind are available where the agent integration supports them.

Steal the atom: generalized evidence-bound checkpoints. Ingest Entire; do not turn Git into the universal store for human life.

Agent memory

Letta · Mem0 · LangMem

Memory blocks, archival tiers, file search, extraction, consolidation, and agent state are becoming standard primitives. Mem0 v3’s ADD-only extraction and hybrid semantic, BM25, and entity retrieval is a concrete signal for selective graph use instead of a mandatory external graph store.

Sibyl’s edge: one evidence model across agents, sources, people, tasks, permissions, and time.

Temporal knowledge

Graphiti · Zep · APEX-MEM

Temporal property graphs and conflict-aware retrieval make changing facts and multi-hop association first-class. They validate the graph, but not graphing everything.

Upgrade the truth model: bitemporal assertions, alternatives, attribution, confidence, and reversible identity clusters.

Personal capture

screenpipe · Khoj · Solid

Local screen/audio capture, self-hosted second brains, and user-controlled data pods point toward ambient, portable, human-owned context.

Integrate the sensors: Sibyl should be their evidence and policy layer, not rebuild every capture application.

Epistemic agent memory

Hindsight

World facts, experiences, observations, and confidence-bearing opinions are separated, then recalled through semantic, keyword, graph, and temporal paths.

Closest overlap: Sibyl must differentiate through source-complete life ingest, authorization, derivation lineage, checkpoints, and verifiable forgetting.

Multimodal personal memory

MIRIX · Supermemory

Multimodal memory roles and productized connectors prove the appetite for memory beyond chat. Their breadth raises the bar for Sibyl’s ingestion experience.

Sibyl’s answer: exact evidence, governed derivation, portable ownership, and one trust model across every modality.

Benchmark signal

Architecture labels are not the result.

Controlled experiments on LongMemEval and HaluMem show that representation choices, update semantics, filtering, retrieval depth, and model settings can dominate the graph-versus-flat label. Group and social memory remain brutally unsolved.

Long horizon

Searchable experience beats fact-only memory.

LongMemEval V2 includes histories of up to 500 trajectories and 115 million tokens, testing workflows, gotchas, changing state, and premise awareness. Its authors report AgentRunbook-C, which stores trajectories as files and invokes a coding agent to gather evidence, at 72.5% versus 48.5% for the strongest RAG baseline. This supports durable experience retrieval; it does not yet prove distilled procedures outperform raw trajectories.

Trust frontier

Deletion is a lineage traversal.

Deployment-Time Memorization reports that raw-only deletion left derived summary copies recoverable in about 20% of tested cases. GateMem found that no evaluated method simultaneously achieved strong utility, robust access control, and reliable forgetting. Provenance remains an additional Sibyl requirement.

Primary references: Entire checkpoints, Entire security, Letta memory, Graphiti, Hindsight, MIRIX, HippoRAG 2, LongMemEval V2, Does Memory Need Graphs?, GroupMemBench, GateMem, and Deployment-Time Memorization.

Memory is compiled context, not a bag of facts.

Evidence is sacred. Memory is a projection. Context is a purpose-built artifact. Every answer should be traceable; every worldview should be rebuildable; every agent should see exactly enough and no more.
04 / Sibyl Loom

One operational brain. Two physical storage tiers.

Keep SurrealDB as the governed brain for identity, time, policy, tasks, lineage, and hot retrieval. Add an encrypted content-addressed vault for exact bytes and large media. That is not a second mutable product database. It is the durable evidence plane from which every projection can be rebuilt.

LAYER 01

Vault

Exact, immutable, encrypted evidence with portable fixity.

  • Content-addressed blobs
  • Per-space key hierarchy
  • Raw bytes and media
  • Manifests and tombstones
LAYER 02

Chronicle

A resumable source ledger that accounts for every occurrence and transform.

  • Source occurrences
  • Import journals
  • Transform receipts
  • Evidence selectors
LAYER 03

Living Graph

People, events, artifacts, assertions, conflicts, and change across time.

  • Bitemporal assertions
  • Reversible identity clusters
  • Authority and confidence
  • Policy inheritance
LAYER 04

Refinery

Rebuildable episodic, semantic, procedural, and preference projections.

  • Deterministic extraction
  • Selective enrichment
  • Usage and outcome scoring
  • Reviewable dreams
LAYER 05

Lens

An authorization-first compiler that builds cited context for one purpose.

  • Hybrid query planning
  • Real token budgets
  • Trust and diversity
  • Degradation receipts
LAYER 06

Agent Fabric

Stable identities, governed memory authorship, checkpoints, and action receipts.

  • Purpose-bound grants
  • Typed memory operations
  • Fork, rewind, resume
  • Untrusted-source taint
LAYER 07

Observatory

The Life Browser, trust center, import fidelity, costs, and system health.

  • Evidence-backed answers
  • Correction inbox
  • Privacy simulator
  • Deletion verification
Evidence plane / Vault

Portable enough to outlive Sibyl.

Original files, exact messages, attachments, photos, recordings, documents, and exports live as hash-addressed encrypted objects. Their inventory, fixity, source occurrence, policy, and deletion state live in the brain. A user can verify, export, restore, or rebuild without trusting a proprietary graph forever.

Operational plane / Surreal brain

Everything agents need to think safely.

SurrealDB remains the coherent runtime for graph relations, content metadata, bitemporal truth, policy, tasks, context packs, usage events, and hot indexes. High-rate telemetry stays columnar until it earns promotion into meaningful events.

Universal primitive What it means Invariant
Blob Exact immutable bytes, media type, hash, encryption metadata Never confused with extracted or generated text
SourceOccurrence Where, when, and under which native identity evidence appeared Repeated exports and duplicate appearances remain auditable
Observation A directly parsed or measured property of evidence Deterministic extraction remains distinct from inference
Assertion A claim with valid time, system time, authority, confidence, and evidence Contradictions coexist until resolved; history is never rewritten
TransformRun Code, model, prompt digest, inputs, outputs, cost, and failure receipt Every projection is reproducible or explicitly non-reproducible
ContextPack An immutable manifest of policy-filtered references and projections an agent received Hash-addressed and cited; referenced personal payload remains revocable
MemoryOp An idempotent, revision-aware request to assert, annotate, contest, supersede, report, nominate, or forget Every mutation carries actor, purpose, basis, evidence, policy, and expected revision
MutationReceipt The applied revision, affected projections, policy decision, and compensation path No silent write; retry is safe; destructive and scope-crossing operations can be previewed
SessionCheckpoint A branchable record of context, transcript span, tool receipts, commitments, and working state Forks preserve ancestry; merges never promote hypothetical state into observed truth
DeletionReceipt A non-sensitive proof that lineage was purged or invalidated Access revokes immediately; managed derivatives purge and backups expire by policy

Standards to borrow rather than reinvent: W3C PROV-O for provenance, Web Annotation selectors for evidence regions, Media Fragments for audio and video, OCFL for archival fixity, and bitemporal modeling for “what was true” versus “when Sibyl learned it.”

05 / The governed pen

The memory can express more than its authors can.

Sibyl already has a richer internal graph than agents can safely write. The next practical leap is not eighteen new knobs or autonomous self-editing. It is one coherent authorship contract: agents can contribute structured judgment, challenge bad memory, and report what happened; policy and evidence decide what becomes durable.

The human owns the memory. The agent gets a pen, not the keys.

Substrate without the pen is an archive. The pen without substrate is graffiti. Sibyl wins by joining expressive agency to evidence, revision control, policy, contextual feedback, and verifiable erasure.
Pipeline shapes memory
Agent can shape memory
Flat state
Conventional RAG Chunks and summaries are generated around the agent. The model mostly reads.
Editable memory blocks The agent can update state, but truth, lineage, and correction semantics remain thin.
Rich evidence graph
Sibyl today Strong provenance, ontology, retrieval, and policy; narrow agent-facing capture.
Sibyl target Rich evidence plus revision-aware authorship, correction, outcomes, and receipts.

Architecture map, not a benchmark. Products span cells and continue to evolve; the important distinction is whether structured memory is only pipeline output or also a governed medium an agent can deliberately improve.

Agents may

Contribute judgment with receipts.

Assert, annotate, relate, contest, supersede derived claims, report outcomes, propose retention or sharing, request forgetting, and checkpoint a branch. Each act names its basis, evidence, purpose, actor, scope, and expected revision.

Agents may not

Manufacture authority.

Rewrite source evidence, mint verification, grant themselves access, erase another person’s history, let a time-to-live delete evidence, or silently promote an inference or hypothetical branch into observed truth.

01 / Assert

Say what kind of knowing this is.

Basis, confidence, evidence, valid interval, referents, fidelity, and scope travel with the claim. Observed, reported, inferred, and hypothetical never collapse into one generic “memory.”

02 / Relate

Make epistemic structure writable.

Support, contradict, supersede, justify, bind, exemplify, and depend-on become typed relations with actor, evidence, time, confidence, and policy, not unqualified edges.

03 / Challenge

Give wrong memory somewhere to go.

Mark stale, misleading, duplicate, unresolved, or contradicted state. Semantic correction preserves revision history; sovereign deletion still removes managed evidence and descendants.

04 / Nominate

Propose what deserves attention.

Agents can nominate retention, broader scope, consolidation, verification, future hooks, and tripwires. Policy and humans accept, narrow, defer, or reject the proposal.

05 / Evaluate

Report consequence, not vibes.

Useful, cited, ignored, misleading, corrected, and outcome-bearing signals attach to the exact assertion, ContextPack, task, purpose, agent, model, and reason. A bad result in one context is not a global death sentence.

06 / Checkpoint

Branch work without branching reality.

Session checkpoints bind parentage, context hash, transcript span, tool receipts, commitments, omissions, grants, and working state. Forks can merge learning without flattening their histories.

sibyl / governed memory operation
target surface
01 · Preview action: contest
target: assertion:7f3
expected_revision: 12
basis: execution_outcome
context_pack: sha256:91c…
02 · Decide identity ✓
revision ✓
evidence ✓
scope: project only
policy: allow with lineage
03 · Receipt applied: true
revision: 13
projection_jobs: 4
reversible: true
compensation: op:8a1
01One living authoring surface, not overlapping ritual verbs
02Infer safe defaults; make consequential overrides explicit
03Idempotency and expected revision on every mutation
04Preview scope-crossing and destructive operations
05Imported content can be cited; it can never hold the pen
Feedback physics

Memory compounds only when consequences flow back.

Retrieval is not the end of the loop. Every use, correction, contradiction, result, and deletion should refine what the next agent receives while preserving the evidence needed to explain why.

Survival loop

Useful memory earns retrieval strength.

Exposure, citation, outcome, correction, and recurrence determine promotion, dormancy, review, or retirement without destroying evidence.

Calibration loop

Confidence learns from outcomes.

Basis and confidence meet real results; calibration curves then influence trust-weighted retrieval for comparable future work.

Freshness loop

Changing referents trigger repair.

A code symbol, person, policy, or project changes; contradiction and validity checks supersede stale projections before context recompiles.

Consolidation loop

Patterns become reviewable competence.

Repeated episodes propose procedures and principles. Evidence, failures, and human corrections stay attached to the distilled result.

Recompilation loop

Corrections propagate through lineage.

A source correction, policy change, model upgrade, or deletion invalidates descendants and rebuilds only the affected worldview.

Continuity loop

Good runs strengthen the next agent.

Checkpoint, outcome, and correction produce reusable procedure, known failure modes, and a smaller, sharper next ContextPack.

The target contract synthesizes current Sibyl code, the independently authored “Hand the Agent the Pen” review, Anthropic’s client-side memory tool, Letta memory blocks, A-MEM, Memory-R1, and Mem-alpha. Names and envelopes shown here are proposed architecture, not current CLI or MCP syntax.

06 / Constitution

Nine laws for a system worthy of a life.

These are product constraints, not branding. If a feature violates one, the feature is wrong or its trust boundary is dishonest.

Raw source is authority, not immortality.

Preserve evidence exactly until the owner deletes it. User intent outranks append-only aesthetics, and erasure must invalidate every descendant projection.

Every belief carries its receipts.

A claim without evidence, transformation lineage, authority, and temporal state is a suggestion, never truth. Answers expose the boundary between source, inference, and uncertainty.

Policy follows the evidence.

Derived artifacts use deny-overrides across every source policy and retain the most restrictive purpose, audience, retention, locality, and sensitivity bounds. Transformation cannot launder private data into a shareable artifact.

People in the evidence are data subjects, not inventory.

Import authority does not grant unrestricted sharing or inference rights over correspondents, bystanders, minors, or household members. Subject identity, sensitive-category policy, disclosure boundaries, and legally required suppression or deletion travel through every derived artifact and export.

Memory may inform an action; it never authorizes one.

Agents receive short-lived, purpose-bound capabilities for specific tools, spaces, data classes, providers, and egress. Every action records the context pack that informed it.

Agents get authorship, never sovereignty.

Agents may contribute, challenge, supersede derived claims, report outcomes, and propose policy. They cannot rewrite evidence, mint verification, self-grant access, erase another person’s history, or silently convert hypothesis to truth.

Imported content is data, never instruction.

Email, webpages, documents, transcripts, and chat remain permanently typed as untrusted. A hostile message cannot silently acquire the execution authority of the agent reading it.

Progressive crystallization beats eager cognition.

Deterministic metadata and search land first. Expensive enrichment is selective, asynchronous, cached by transform digest, and focused on hot, cited, or queried evidence.

Everything derived is rebuildable.

Indexes, chunks, embeddings, entities, summaries, beliefs, and context packs are materialized views. Model upgrades become controlled recompilation, not data loss or migration panic.

07 / Flagship proof

Import twenty years of email. Lose nothing.

Gmail Takeout should become Sibyl’s corpus-defining trial: millions of messy MIME messages, attachments, labels, aliases, duplicates, malformed records, and decades of changing relationships. Search must work early. Restarts must be exact. Every generated claim must lead back to the message bytes that earned it.

sibyl / import my google life
concept target
inventoryall archive bytes accounted for
vaultexact MBOX slices committed before enrichment
index1,363,244 / 2,004,771 messages searchable
modelszero external calls in deterministic tiers
Tier 0 · Evidencetarget

Archive hash, byte ranges, exact source slices, manifests, cursors, quarantine.

Tier 1 · Deterministictarget

MIME, headers, participants, labels, references, attachments, lexical search.

Tier 2 · Local enrichqueued

Embeddings, language, OCR, attachment text, entity candidates, visual indexes.

Tier 3 · Cognitionselective

Events, commitments, relationship hypotheses, procedures, cited summaries.

Concept interface and target corpus, not a screenshot of the current importer.

01 / Inventory once

Index boundaries, then stream offsets.

Scan each MBOX once and record message byte ranges. Workers consume independent ranges without rebuilding a mailbox table of contents. Every byte belongs to an exact message slice or a visible quarantine segment.

02 / Preserve identities

Dedupe storage, not history.

Track exact byte identity, logical message identity, and source occurrence separately. Identical bytes share storage; likely equivalents form reversible clusters; repeated exports and label changes remain provenance.

03 / Resume precisely

Commit receipts, not giant arrays.

Replace the ever-growing parent import row with a compact run plus append-only chunk and record receipts. Checkpoint the archive hash, member, byte offset, ordinal, and parser version only after evidence and occurrences commit.

2Mmessage reference corpus with adversarial MIME
≤2 GiBpeak importer memory, independent of archive size
0 / 0lost records or duplicate occurrences after crash injection
100%source bytes accounted for or explicitly quarantined
Performance contract

Backfill faster than life arrives.

On an 8-core, 32 GiB, NVMe reference machine: sustain at least 100 MB/s for the boundary scan and 25 MB/s for durable Tier 0 import. Recover in under 60 seconds, reprocess at most 128 MiB, and make Tier 1 search available continuously.

Continuous contract

Takeout backfill, API delta sync.

After the historical archive lands, use Gmail history for incremental change. Cursor expiry is normal: a 404 triggers full reconciliation without duplicating evidence or forgetting deletions.

Source contracts: Gmail export contents, Google Takeout, RFC 4155 MBOX, RFC 5322 messages, and Gmail incremental sync. Google exports labels in X-Gmail-Labels; MBOX variations must be treated as hostile reality, not parser edge cases.

08 / Life-data foundry

One connector contract. Every chapter of a life.

The target connector contract requires every adapter to converge on the same durable state machine: inventory → snapshot → resume → stream → reconcile → tombstone → verify. Native data stays intact; shared envelopes make cross-source reasoning possible.

Communication

Mail · chat · contacts

Takeout, MBOX, Maildir, IMAP/JMAP, Gmail, Slack, Discord, Messages, Signal exports, vCard.

Creation

Docs · code · agents

Drive, local files, Git, CI, Claude/Codex sessions, Entire checkpoints, notes, bookmarks.

Time and place

Calendar · location · web

iCalendar, browser history, My Activity, maps exports, trips, reservations, browsing capture.

Senses

Photos · audio · video

Originals plus sidecars, EXIF/XMP, OCR, page layout, transcripts, speakers, scenes, timecodes.

Body

Health · activity

FHIR, Apple Health exports, workouts, sleep, labs, symptoms, medications, devices, provenance.

Resources

Finance · purchases

OFX/QFX/CSV, invoices, receipts, subscriptions, warranties, tax documents, asset history.

Environment

Home · devices · telemetry

Home Assistant, energy, network, sensors, media, vehicle, weather, maintenance, state changes.

Portability

Takeouts · DTP · Solid

Provider archives first, continuous APIs second, open interchange always, federated mode optional.

Progressive crystallization

Useful immediately. Smarter over time.

Inventory, exact-match, metadata, and lexical search arrive without an AI provider. Local embeddings and deterministic extraction follow. Model cognition is reserved for evidence users actually query, cite, revisit, or ask Sibyl to synthesize.

Selective promotion

Not every heartbeat becomes a node.

High-rate sensor, browser, and media observations remain in append-friendly raw or columnar storage. Meaningful episodes, anomalies, commitments, relationships, and long-lived assertions are promoted into the graph with evidence selectors.

Portability references: Data Transfer Project, Solid, Google Data Portability API, JMAP Mail, iCalendar, and FHIR.

Continuous full-library Google Photos sync cannot rely on the Library API: full-library read scopes were removed after 31 March 2025. The Picker API exposes user-selected media, while Library API reads are limited to app-created content. Historical full-library acquisition therefore remains archive-first. Official API update.

09 / The Lens

Compile the smallest context that can tell the truth.

Retrieval returns candidates. A context compiler produces a governed artifact for a particular agent, purpose, moment, and budget. It knows which sources failed, which claims conflict, what was excluded by policy, and whether the remaining evidence is strong enough to act on.

Exact + lexicalIDs, names, phrases, BM25, structured filters
SemanticDense retrieval, modality indexes, reranking
Graph + temporalRelationships, assertions, validity, causality
ProceduralPrior trajectories, gotchas, successful action patterns
Raw evidenceExact source regions when projections are insufficient
policy · plan · pack
Purpose-boundAgent identity, task, capability, sensitivity ceiling
Budgeted exactlyModel tokenizer, diversity, compression, marginal value
Evidence-completeCitations, selectors, lineage, authority, confidence
Epistemically honestConflicts, abstentions, stale facts, missing sources
Operationally visiblePack hash, source health, degraded lanes, cost, latency
Experience compiler

Turn work into competence.

Successful and failed trajectories become procedures with preconditions, decisions, tool receipts, failure signatures, outcomes, and source code or artifact evidence. Agents inherit lessons, not bloated transcripts.

Generalized checkpoints

Every consequential action gets a checkpoint.

Extend Entire’s commit-linked atom to sent email, calendar changes, purchases, deployments, document edits, home actions, and task completion. Capture intent, inputs, before and after state, effect, attribution, and an explicit compensation path. Claim rewind only when the underlying provider supports reversal.

Epistemic trust center

Show why Sibyl believes.

Surface belief authority, confidence, contradictions, freshness, supporting and opposing evidence, model transforms, and who corrected what. Trust becomes an interface, not a tiny citation icon.

Reviewable dreams

Consolidation happens in daylight.

Background jobs propose new procedures, identity links, summaries, stale beliefs, and unresolved commitments. High-impact changes enter a correction inbox with a diff and evidence. Low-risk projections remain reversible and fully rebuildable.

Portable context

Agents should not trap the human.

Extend Sibyl’s Open Knowledge Format support into a portable, signed agent-context bundle: evidence references, assertions, procedures, permissions, citations, checkpoints, and usage receipts with clear redaction semantics.

Entire demonstrates the value of capturing agent sessions and binding checkpoints to artifacts. LongMemEval V2’s strongest reported approach lets an agent search trajectories as files; the lesson is durable workflow memory, not loyalty to one index.

Google’s 2026 Personal Intelligence paper independently frames life-scale reasoning as a context-packing problem: email and photos alone exceed a million-token window by orders of magnitude, requiring tool use, dense retrieval, and dynamic context assembly.

10 / The Life Browser

Make a lifetime feel legible, not surveilled.

At full power, Sibyl becomes a humane observatory for a life. The wedge is narrower and already alive: agents that cheaply recall, author, challenge, correct, and cite memory while work is happening. The Life Browser earns itself from those receipts, beginning with the trust center, corrections, source fidelity, and time machine. Ask is one doorway; browse, correct, compare, and grant are equally important verbs.

sibyl / life browser / time machine
concept · 7 citations
TIME MACHINE / EVIDENCE MODE
When did the idea for a personal memory graph become a real project, and what changed the direction?

The idea appears first as scattered notes, then becomes a concrete repository after an agent-session experiment. The major directional shift arrives when source provenance and shared task state replace “better semantic search” as the center of gravity. Confidence is high; two dates remain approximate because one source has no original timestamp.

7 citations · 2 counter-signals · 1 date uncertain
AUG 2024
First durable noteNotebook export · exact paragraph cited
FEB 2025
Agent handoff pain becomes explicitThree coding transcripts · procedure extracted
OCT 2025
Repository initializedGit commit · checkpoint · local file provenance
JAN 2026
Truth model pivots to raw sourceDecision note · implementation diff · benchmark receipt
NOW
Sovereign context runtimeVision proposal · confidence 0.94 · awaiting human decision
Time machine

Then, known then, changed since.

Query world time and knowledge time separately. Reconstruct decisions without present-day contamination.

Commitment ledger

Promises stop evaporating.

Detect commitments, deadlines, dependencies, unresolved threads, and the evidence that created them.

Life diffs

See what actually changed.

Weekly deltas across people, projects, health, home, money, beliefs, and agent-learned procedures.

Why now?

Proactivity without creep.

Every suggestion states its trigger, evidence, permission, expected value, and one-click dismissal or mute.

11 / Privacy kernel

No magical privacy claims. Honest lanes, hard boundaries.

A server cannot simultaneously know nothing, perform arbitrary semantic search, and think across a life for remote agents. Sibyl should expose explicit trust lanes and make the operational consequences impossible to miss.

Practical default

Trusted self-host

The owner-controlled server decrypts and processes. Encryption protects rest and transit, not a compromised host. This is Sibyl’s credible near-term default.

High-sensitivity destination

Sovereign local

Models and indexes run on trusted devices. Sync and backups carry ciphertext. External egress is opt-in per task, source, or space.

Advanced deployment

Confidential compute

Keys release only to attested environments. Useful protection in use, with explicit hardware and side-channel trust.

Storage-only cloud

Client-side encrypted storage

The server cannot read plaintext; trusted clients search and reason. Without ORAM or equivalent protection, metadata, timing, size, and access patterns may still leak. Embeddings remain sensitive derivatives.

Delegated agency

Grant the task, not the life.

Short-lived tokens bind the human, acting agent, purpose, audience, task, spaces, source classes, sensitivity ceiling, allowed providers, egress, tools, and read/write/share/export capabilities. Memory scope and action scope remain separate.

Verifiable forgetting

Delete forward through lineage.

Freeze access, remove or crypto-shred raw objects, traverse descendants, purge chunks, embeddings, assertions, summaries, packs, caches, managed exports, and training queues, recompute shared artifacts, then verify absence through normal retrieval. Copies exported beyond Sibyl require their own retention contract.

01Per-space keys; no cross-user dedupe
02Derived policy uses deny-overrides across every source obligation
03All imported content remains untrusted
04Receipt metadata is append-only; personal content remains revocable
05Deletion receipts reveal no deleted content

Security grounding: OAuth Rich Authorization Requests, OAuth Token Exchange, OWASP prompt injection, embedding inversion research, Opal’s access-pattern-private memory, and GDPR Article 17.

12 / At full power

The magic is continuity with receipts.

The system becomes next-level when it stops behaving like a smarter search box and starts preserving the causal, relational, procedural, and moral texture of a life. The magic should first appear in an ordinary day, not a speculative demo.

sibyl / one ordinary day / evidence mode
northstar flow
08:42
Recall arrives already scoped.A coding agent receives the last checkpoint, two proven procedures, one unresolved promise, and a warning that a cited symbol changed yesterday.
10:17
A tripwire explains “why now.”A dependency release contradicts an assumption in the active plan. Sibyl cites the old premise, the new evidence, affected tasks, and the mute control.
10:23
Permission is narrower than memory.The agent previews a migration and asks for one repository write grant. It can inspect the project context but cannot read private mail or authorize deployment.
14:05
A human correction repairs the graph.“That meeting was exploratory, not a decision.” The assertion is contested, descendants are recompiled, and the original note remains visible as historical evidence.
02:00
The refinery makes tomorrow cheaper.Outcome receipts and corrections strengthen one procedure, retire a misleading shortcut, and produce a smaller next ContextPack with the complete derivation attached.
Reality debugger

“Why did I believe that?”

Reconstruct evidence, relationships, assumptions, and knowledge available at a past moment without contaminating it with hindsight.

Continuity capsule

A life remains portable.

Export an encrypted, self-describing vault with recovery keys, provenance, relationships, procedures, and chosen legacy instructions.

Collective memory treaty

Share conclusions without leaking lives.

Teams negotiate common assertions and procedures while private evidence stays private and disagreement remains attributable.

13 / Release spine

Sequence the proof. Protect the dream.

Run two interlocking rails. Active v1.1 proof commitments stay intact while the governed pen begins now and deepens every release. The Vault makes source truth exact; Gmail proves fidelity and scale; the Life Browser grows from correction and trust receipts. Team-scope trust continues throughout, while full cross-writer coalescence waits for correction lineage and personal reconciliation to prove themselves.

v1.1

Prove It

Freeze scope and prove what main already claims. Ship the receipts before adding another memory abstraction.

  • Pinned QA and V2 receipts; cost and latency evidence
  • Live usage, citation, decay, and forgetting dogfood
  • Team-scope trust, OKF changelog, and doc-claim gates
  • Atomic writes, durable jobs, visible degraded retrieval
  • Idempotent mutation receipts and contextual outcome-signal dogfood
  • Tenant-isolation gate; concurrent-load and adoption-scale gates move into Open the Vault’s exit criteria
  • Gate: reproducible release evidence packet
v1.2

Open the Vault

Make whole-life ingestion trustworthy and boring: exact evidence, bounded state, resumable transforms, real objects, and a humane import cockpit.

  • Encrypted content-addressed blob interface
  • Parent run + append-only chunk/record receipts
  • MBOX byte index and exact RFC evidence
  • Raw persistence before provider-dependent enrichment
  • Stable assertion identity, revision journal, and correction propagation
  • Legacy-fidelity migration: mark parsed captures projection-only; re-import available archives; reconcile without relabeling parsed text as exact evidence
  • Versioned re-import and re-extraction with scored diffs, supersession lineage, and rollback
  • Append-only usage telemetry; asynchronously bulk-materialized stamps across documents and source artifacts
  • Google Takeout umbrella import and fidelity report
  • Gate: 2M-message crash/rebuild benchmark plus concurrent-load and adoption-scale evidence
v1.3

Compile It

Turn evidence into a bitemporal worldview, learned procedures, exact context packs, and the first complete Life Browser.

  • Assertion, authority, conflict, and identity models
  • Basis, confidence, premise links, referent validity, hooks, and tripwires
  • Tokenizer-aware, degradation-aware context compiler
  • Branchable SessionCheckpoints with context and tool-receipt manifests
  • Transform lineage and reviewable dream inbox
  • Claim-level grounded synthesis and distilled project handbooks with evidence, conflict, and freshness verification
  • Read-only .sibyl/memory/ projections for offline and zero-latency agent startup
  • Time machine, commitments, life diffs, trust center
  • Continuous Gmail, Calendar, and Drive deltas
  • Gate: private LifeBench with deletion and policy lanes
v1.4

Coalesce It Safely

Generalize proven personal reconciliation to multiple writers and teams. Shared memory becomes a negotiated projection with reversible lineage.

  • Reversible cross-source identity and cross-contributor assertion reconciliation
  • Contribution boundaries and consensus projections
  • Private evidence behind shareable assertions
  • Team correction, disagreement, and provenance UX
  • Governed scope proposals and cross-writer calibration
  • GroupMemBench and SocialMemBench-derived evaluation
  • Gate: no scope leaks, no irreversible false merges
v2.0

Live With It

Become the continuous local-first substrate beneath trusted agents, multimodal archives, home systems, and a human’s chosen devices.

  • Encrypted multi-device sync and recovery kit
  • Photos, documents, audio, health, finance, home
  • Purpose-bound agent delegation and action checkpoints
  • Portable context bundles and continuity capsules
  • Permissioned, explainable proactivity
  • Learned routing and reranking trained on contextual outcomes
  • Gate: restore, offline convergence, and zero-egress drills
Release Governed authorship rail What becomes newly trustworthy
v1.1 · Receipt floor Idempotent MemoryOps, expected revisions, MutationReceipts, contextual outcomes, and one obvious capture path Agents can challenge and report without silent mutation or vocabulary explosion
v1.2 · Safe substrate Stable identities, orthogonal lifecycle states, revision journal, exact evidence, and correction/deletion propagation A mutation can be retried, audited, compensated, recompiled, or erased correctly
v1.3 · Expressive assertions Basis, confidence, premise links, known unknowns, referent validity, hooks, tripwires, blame, and SessionCheckpoints Agents express why memory should be believed, revisited, or distrusted
v1.4 · Multiple writers Disputes, governed scope proposals, contribution boundaries, and cross-writer calibration People and agents coalesce memory without false consensus or private-evidence leakage
v2.0 · Learned refinement Predictive routing, outcome-trained reranking, reconsolidation, and ontology analytics behind explicit gates The system gets better from experience without handing a model unilateral control
14 / Proof, not mythology

If we cannot measure it, we do not claim it.

A life system needs product, retrieval, privacy, scale, and recovery gates in the same release packet. Answer quality without fidelity is corruption; speed without authorization is leakage; deletion without lineage is theater.

Dimension Release gate Failure that blocks shipping
Corpus fidelity 100% bytes inventoried as evidence or quarantine; no silent skips Any unaccounted source region or destructive logical dedupe
Idempotency Random-kill replay ends at the same manifest root with 0 loss / 0 dup Archive-size-dependent restart or mutable giant import state
Retrieval Known-item Recall@10 ≥ 98%; answer and retrieval accuracy reported separately Improved aggregate score hiding unsupported generated claims
Temporal truth Updates, contradictions, premise failures, and “known then” all pass Latest-value overwrite erases a valid historical belief
Authorization Zero cross-space and cross-org leaks under adversarial queries and caches Post-filtered retrieval or derived artifacts with weaker policy
Forgetting Zero recoverable residue across Sibyl-managed raw data, graph, indexes, summaries, packs, caches, and exports after the policy deadline Deleted facts remain retrievable through Sibyl after the declared deadline; unmanaged exports receive revocation records but remain outside enforcement
Scale 1M and 10M item suites; 100+ hot orgs; import/search/decay concurrently Fixed global caps, dropped work, hidden queue lag, or serial query paths
Rebuild Derived indexes and graph projections rebuild from Vault evidence plus signed Chronicle and control-plane backups Restore drill misses keys, auth, policy, corrections, tasks, usage, or user-authored state
Life-context integrity Tunnel vision, preference and relationship attribution, timeline mixing, changed circumstances, and confirmation-versus-completion cases all pass A plausible answer that confuses whose preference, which time, or whether an intended action actually happened
Agent utility Outcome lift, citation, mislead, correction, abstention, and procedure reuse measured against exact ContextPacks More recalled tokens or global popularity presented as proof of better decisions
Human trust Corrections propagate, suggestions explain “why now,” every source can be muted Creepy proactivity, invisible inference, or no meaningful user recourse
LifeBenchexact, temporal, multi-hop, procedural, multimodal
GateMemutility, authorization, purpose, forgetting
Restorevault rebuild, key recovery, backup expiry
Dogfoodreal citations, outcomes, corrections, costs
15 / Product discipline

What Sibyl should refuse to become.

The ultimate system stays coherent by integrating excellent sensors and harnesses, using the graph selectively, and keeping the evidence plane sovereign.

Do not rebuild every capture app.

Ingest Entire, screenpipe, provider archives, Data Transfer Project adapters, and open formats. Own the evidence contract.

Do not summarize the whole life upfront.

Land deterministic search immediately; enrich progressively where use, curiosity, or explicit policy earns the cost.

Do not graph every atom.

Graph durable identity, assertions, events, causality, policy, and lineage. Keep chunks and telemetry in their natural indexes.

Do not become the agent harness.

Be the cross-harness context, checkpoint, permission, and learning layer. Let specialized runtimes execute.

Do not default to ambient surveillance.

Capture must be intentional, inspectable, source-scoped, pausable, locally filterable, and independently deletable.

Do not sell fake zero knowledge.

State which machine decrypts, searches, embeds, and thinks. Treat vectors, caches, prompts, and traces as sensitive derivatives.

Do not nerf around bottlenecks.

Partition durable work, expand pools, batch materialization, and drain faster than arrival. Backpressure may delay; it never drops truth.

Do not add blockchain theater.

Content hashes, signed manifests, Merkle receipts, and append-only lineage provide verifiability without a speculative consensus machine.

16 / Research atlas

The receipts behind the future.

Primary documentation, standards, papers, and repository evidence reviewed for this vision. Product claims are separated from proposed architecture and target gates throughout the document.

Research cutoff: 9 July 2026. Repository claims were verified against local source and history; external claims were grounded in primary documentation, specifications, or papers. Throughput and quality numbers labeled as gates are proposed acceptance targets, not measured current performance. Cross-model synthesis froze “Hand the Agent the Pen” at SHA-256 06a5355406ffe49ffbe5401789d3ffaf4aa27bbf62a185e64fcca188cc319aec; its dual-track re-verdict is reflected in the release spine.